orthogonal projections forming 2--D scatterplots. An important task in exploring a class
structure in higher dimensions is to understand what discriminates individual classes.
Unfortunately, as the number of dimensions of the data spaces increase, or the number of
distinct classes increase, mapping from data space into scatterplots often involves a loss
of separation among individual classes; or even worse, it can show a misleading picture of
a class structure well defined in higher dimensions. This leads to a major challenge in
real world visual data analysis: What are promising strategies that let classes visually stand out
more clearly in 2--D orthogonal embeddings?
In my talk I will discuss a novel way to make individual classes more visible when the
original embedding of a class structure into a given 2--D scatterplot is mixing. The
basic idea is to find a visually more effective embedding of a subset of the classes
satisfying a tolerable amount of separation violation. We call such embeddings consistent
class embeddings. One difficulty in real world scenarios is that most of the
scatterplots show a misleading picture of a well-defined class structure in n--D.
I will demonstrate how we applied consistent class embeddings to support the human
analyst in exploring large SPLOM's.